6409 - OOP-FP GPU Programming and Debugging using Open-Source GpuScript.pdf

1、Alan RockOOP/FP GPU Programming and Debugging using Open-Source GpuScript OOP/FP GPU Programming and Debugging using Open-Source GpuScript Alan RockARTIFICIAL INTELLIGENCE(AI)HardwareSingle GPU in a single PC or laptop computerSupercomputersGPU super-clustersGPU Development ToolsWrite,modify,and deb

2、ug GPU codeCritical routines to entire applicationsAlgorithms combined with programming techniquesReduce algorithm time complexity from O(N!?)to O(1)Solve problems in seconds instead of years or decadesEnhanced High-Performance ComputingThe picture cant be displayed.HardwareBuildPower/hrTime Share/h

3、rSpeed,TFLOPSCPU$1,000$0.01$0.01 0.1GPU$1,500$0.01$0.01 20GPU with GS$1,500$0.01$0.01 20,000Supercomputer$100,000,000$1,000$5,000 800,000GPU supercluster$100,000,000,000$25,000$160,000 260,000,000The picture cant be displayed.Highly productive developmentFull debugging supportCode generationExtensiv

4、e libraries for both graphics and computationEasy to learn and useFamiliar development environmentFamiliar high-level programming languageWidely accessible,free and open sourceGPU Development ToolsParallel algorithms can reduce an algorithms time complexity,or Big-O,within limits of GPU memory and p

5、rocessor count.FFT=2048 samplesMatrix multiply=2048x2048Traveling Salesman=4500 citiesParallel AlgorithmsAlgorithmCPUGPUMinO(n)O(1)MaxO(n)O(1)SumO(n)O(1)SortO(n log n)O(1)FFTO(n log n)O(1)Matrix multiplicationO(n3)O(1)Matrix inversionO(n3)O(1)Singular Value Decomposition O(n3)O(1)Dynamic Programming

6、O(2n)O(n)Traveling SalesmanO(n!)O(n)Neural NetworksO(n!)O(1)Algorithms:Billion Nested Loop100,000 times faster:1 sec:1 dayMatrix MultiplicationCPU:O(N3)GPU:O(N)GPU with atomic functions:O(1)Prime Numbers:Sieve of EratosthenesGPU:O(1)Algorithms:Counting sort is O(1)for 2048 lists